For fast measurements of single seeds using near infrared (NIR) spectra for the prediction of seed moisture content, it may be necessary to reduce the spectra from over 1000 wavelengths to just a few narrow bands. This reduction makes it possible to utilise a few parallel and simultaneous NIR sensor measurements in seed sorting instead of scanning a few NIR bands that are sequential in time. Three different approaches of genetic algorithms (GA) were used to select wavelengths within the range 400–2500 nm. The GA models were compared for two different datasets: single seeds and bulk seeds of Scots pine. It was shown that GA and interval partial least squares combined with GA allowed a meaningful reduction in spectral content without any loss in predictive quality. The three models selected three to six wavelength regions mainly around the peak of the combination of the first O–H stretching overtone and the O–H bending at 1190 nm and on the slopes of the first O–H stretching overtone at 1450 nm. For some of the GA models, the selected regions were subdivided into one to three more regions. In total six to eight narrow regions were used to simulate uniform density filters based on average absorbance within selected regions. The RMSEP values of the filter simulations were of at least the same quality as those for the whole wavelength range or the NIR range. The wavelength bands chosen for the single seeds were also applied for the bulk samples and vice versa with good result. The overall results illustrate the possibility of using GAs to select wavelengths in order to build filter spectrometers based on a few wavelength bands for the determination of seed moisture content.

Keywords: PLS calibration, PLS prediction, genetic algorithm, calibration set, test set, single seeds, Scots pine, Pinus sylvestris